Evaluating Different Artificial Neural Network Forecasting Approaches for Optimizing District Heating Network Operation

This paper examines the effectiveness of Artificial Neural Networks (ANNs) in enhancing the operation of district heating networks (DHN), which face complexities due to varying conditions like weather, user behavior and energy availability. Accurate heat demand forecasting is crucial for energy-efficient management. We investigate various ANNs, including a combination with Linear Regression, feedforward ANNs, Long Short-Term Memory Networks (LSTMs) and Convolutional Neural Networks (CNNs), as well as the recent Temporal Fusion Transformer (TFT). These methods are evaluated for their ability to predict diverse profiles of heat sources and sinks, aiming for a good generalization capability. Additionally, we focus on providing operators with easily interpretable forecasts together with optimization strategies, emphasizing the importance of comprehensible confidence intervals. The study utilizes two years data from the Stiftung Liebenau DHN, which incorporates multiple energy sources (e.g. CHP, biomass, natural gas) and different heat sinks, e.g., residential buildings, workshops and green houses with highly varying heat demand profiles. Our findings confirm that advanced models outperform simpler ones, effectively capturing trends, though predicting the irregular peaks remains challenging. Economic analysis of the reference DHN suggests that applying predictive methods to forecast heat demand enhances energy efficiency and is economically beneficial, given the low investment cost.